verifylisted

# verify Turn "it looks done" into a verdict backed by executed evidence. All paths below are relative to this skill's directory. ## Three laws 1. **No verdict without an executed command.** Cite the command, exit code, and a line of output. A gate that could not run is *not run*, never passing. 2. **No finding without a refutation attempt.** Every model-authored candidate faces an independent skeptic. Survivors are reported; the rest are counted. 3. **No repair that only silences the checker.** The richer tiers refuse skipped or weakened tests, suppressions, widened types, swallowed errors, edited CI, and rewritten plans. `scripts/forbidden-repairs.mjs` enforces this. ## Modes and tiers Syntax: Codex uses `$verify`; the Claude plugin uses `/maxgfr:verify`; a standalone Claude skill uses `/verify`. | Invocation | Agents | What it does | |---|---|---| | no arguments / `ultralight` | at most 1 | **Default.** Run every detected gate once; no analysis or repair. | | `light` | ~7 + 1/candidate | Gates, plan conformance, 3-lens defect hunt, one skeptic per claim, repair loop. | | `normal` | ~9 + 1–3/candidate | `light` plus behaviour proof and panels on blockers. | | `deep` | ~13 + 1–3/candidate | Every lens, red-green audit, panels throughout. | | `report` | — | Read-only, any tier. Follow with "fix" to apply blockers once. | | `crosscheck` | +1 + 1/candidate | Add a second opinion from the other CLI agent. | | `<ref>` | — | Explicit fixed point (`main`, a SHA, `HEAD~3`). |